Speaker adaptation using codebook integrated deep neural networks for speech enhancement

Deep neural network (DNN) based speech enhancement techniques have shown superior performance compared to the traditional speech enhancement approaches in handling nonstationary noise. However, their performance is often compromised as a result of mismatch between their testing and training conditions. In this work, a codebook integrated deep neural network (CI-DNN) approach is introduced for speech enhancement, which mitigates this mismatch by employing existing speaker adapted codebooks with a DNN. The proposed CI-DNN demonstrates better speech enhancement performance compared to the corresponding speaker independent DNNs. The CI-DNN approach essentially involves a post processing operation for DNN and, hence, is applicable to any DNN architecture.